Fig. 2: Training procedure and schematic description of the machine learning pipeline for the individualized diagnostics.

A Example of one trial from the lexical categorization training. First, a fixation cross is presented, followed by a letter string that is presented until a button press, but for a maximum of 10 s. After a response, feedback is given. A red square indicates an incorrect response (either a word was categorized as a nonword or vice versa), and a green square indicates a correct response. Participants pressed the “f” key for a word categorization and the “j” key for a nonword categorization. B The training design comprises three sessions and a pre-post diagnostic based on a standardized reading speed test (adult version of the Salzburger Lesescreening - SLS; ref. ⁷¹). The pre-test and the first session of the lexical categorization training (dashed box) were used to predict the outcome (dotted box). In Experiments 2 and 3, the same procedure was implemented for control training procedures (Phonics and an adapted lexical categorization training). In Experiment 1, the post-test was conducted after the last training session on day 3. C Analysis steps in the applied cross-validation for the training procedure (upper row) and testing procedure (lower row). The grid on the left represents the datasets. We used a consensus nested leave-one-out-cross-validation; thus, for training, we used all but one dataset to generate features, select features, and train a prediction model. We then applied the trained pipeline to the one left-out dataset for testing. Note we used this cross-validation procedure to prevent extensive overfitting (see methods and supplement for more details).